Gemini thinking_config Not Supported on Model — Fix (2026) | AI Error Hub

GeminiThinkingSeverity: LowHTTP 400

Gemini thinking_config Not Supported on Model

You passed thinking_config to a model that doesn't support it. Only Gemini 2.5 series supports thinking (Pro always-on, Flash configurable, Flash-Lite configurable). Legacy 2.0 Flash, 1.5 Pro/Flash, and the experimental 2.0-flash-thinking-exp preview all reject it. Fix by detecting model support before passing config, or migrating to 2.5.

Error code: INVALID_ARGUMENTHTTP: 400Category: ThinkingLast tested: Nov 15, 2026

Quick Fix (TL;DR)

Check model name before adding `thinking_config`. Only `gemini-2.5-*` models accept it. For older models (`gemini-2.0-flash`, `gemini-1.5-pro`), omit the field entirely. Best practice: use a helper that strips `thinking_config` when the target model is not 2.5 series.

The Error Messages You're Seeing

Response body variants

HTTP/1.1 400 Bad Request

{
  "error": {
    "code": 400,
    "message": "Field 'generation_config.thinking_config' is 
    not supported for model 'gemini-2.0-flash'.",
    "status": "INVALID_ARGUMENT"
  }
}

Variants:
"thinking_budget field only available on Gemini 2.5 models"
"include_thoughts not supported on this model version"
"Model 'gemini-2.0-flash-thinking-exp' has been deprecated. 
Use gemini-2.5-flash instead."
"Unknown field: thinking_config"

Thinking Support Matrix

Model	thinking_config?	Migration path
gemini-2.5-pro	Yes (always on)	Already supported
gemini-2.5-flash	Yes (configurable)	Already supported
gemini-2.5-flash-lite	Yes (opt-in)	Already supported
gemini-2.0-flash	No	Move to 2.5-flash
gemini-2.0-flash-lite	No	Move to 2.5-flash-lite
gemini-2.0-flash-thinking-exp	Preview (deprecated)	Move to 2.5-flash
gemini-1.5-pro	No	Move to 2.5-pro
gemini-1.5-flash	No	Move to 2.5-flash

What Actually Causes This Error

36%

Shared config object across modelsApp uses same config for both 2.5-flash and 2.0-flash routing.

24%

Copy-paste from 2.5 docs to 1.5 codeDocs assume 2.5; older codebases still on 1.5.

18%

Deprecated experimental previewgemini-2.0-flash-thinking-exp retired; migrate.

12%

Environment variable / config driftProd overrides model to older version, config stays.

10%

A/B test with mixed modelsRouting 50/50 between generations without conditional config.

Fixes That Work (Tested Nov 2026)

1Detect Model Capability Before Config

Python · feature-aware config builder
from google import genai
from google.genai import types

client = genai.Client()

def supports_thinking(model):
    """True if model accepts thinking_config field."""
    return model.startswith("gemini-2.5")

def build_config(model, temperature=0.7, thinking_budget=-1):
    """Build config only including supported fields."""
    kwargs = {"temperature": temperature}
    if supports_thinking(model):
        kwargs["thinking_config"] = types.ThinkingConfig(
            thinking_budget=thinking_budget
        )
    return types.GenerateContentConfig(**kwargs)

# Works for both 2.5 and 2.0 models
for model in ["gemini-2.5-flash", "gemini-2.0-flash"]:
    response = client.models.generate_content(
        model=model,
        contents="Explain quantum entanglement.",
        config=build_config(model)
    )
    print(f"{model}: {response.text[:100]}")

2Strip Unsupported Fields Automatically

Python · sanitize config for any model
MODEL_FEATURES = {
    "gemini-2.5-pro":        {"thinking": True, "caching": True, "vision": True},
    "gemini-2.5-flash":      {"thinking": True, "caching": True, "vision": True},
    "gemini-2.5-flash-lite": {"thinking": True, "caching": True, "vision": True},
    "gemini-2.0-flash":      {"thinking": False, "caching": True, "vision": True},
    "gemini-2.0-flash-lite": {"thinking": False, "caching": True, "vision": True},
    "gemini-1.5-pro":        {"thinking": False, "caching": True, "vision": True},
    "gemini-1.5-flash":      {"thinking": False, "caching": True, "vision": True},
}

def sanitize_config(model, config_dict):
    """Remove fields the target model doesn't support."""
    # Match on prefix to catch versioned names like gemini-2.5-pro-002
    features = None
    for prefix, feat in MODEL_FEATURES.items():
        if model.startswith(prefix):
            features = feat
            break
    if features is None:
        return config_dict  # unknown model, pass through
    
    clean = dict(config_dict)
    if not features["thinking"]:
        clean.pop("thinking_config", None)
    return clean

# Usage
raw = {
    "temperature": 0.7,
    "thinking_config": types.ThinkingConfig(thinking_budget=-1),
    "max_output_tokens": 4096,
}
clean = sanitize_config("gemini-2.0-flash", raw)
config = types.GenerateContentConfig(**clean)

Deprecation of experimental thinking preview

gemini-2.0-flash-thinking-exp was Google's preview of thinking on 2.0 Flash. It's deprecated. Migrate to gemini-2.5-flash, which has stable production thinking support and better performance.

3Migration Path from 1.5 / 2.0 to 2.5

Python · gradual rollout
import os, random

# Feature flag driven migration
MIGRATION_PCT = int(os.environ.get("GEMINI_25_MIGRATION_PCT", "0"))

def choose_model(base_model):
    """Route to 2.5 for MIGRATION_PCT% of calls."""
    if random.randint(1, 100) <= MIGRATION_PCT:
        upgrade = {
            "gemini-1.5-pro":   "gemini-2.5-pro",
            "gemini-1.5-flash": "gemini-2.5-flash",
            "gemini-2.0-flash": "gemini-2.5-flash",
        }
        return upgrade.get(base_model, base_model)
    return base_model

# Then use build_config from Fix 1
# Roll from 0% → 10% → 50% → 100% while monitoring quality

Preventing This Error Going Forward

Build config dynamically per model, not statically shared.
Maintain model feature matrix in one place. Single source of truth.
Use prefix-match for versioned model IDs. Match gemini-2.5-pro-002 too.
Prefer 2.5 series for new code. Wider feature support, better quality.
Watch deprecation notices in Vertex AI release notes.
Never mix model gens without config-per-model logic.

Tested by Ahmed R.
Researcher · AI Error Hub

Last verified: Nov 15, 2026 · SDK: google-genai 0.8+

Frequently Asked Questions

Not natively — 2.0 Flash doesn't have the reasoning mode. Best you can do is prompt engineering ("Think step by step, then answer.") which is different from the model-native reasoning tokens in 2.5.

API-compatible yes; behavior differs — 2.5 has thinking on by default (dynamic), slightly higher latency, better quality. Set thinking_budget=0 on 2.5-flash to match 2.0-flash latency profile.

Currently maintained but not receiving new features. Google typically deprecates prior gen ~12 months after next-gen GA. Plan migration before deprecation notice; upgrade paths get tighter as dates approach.

Experimental IDs come and go. Don't rely on them in production. If you must use one, wrap in try/except and have a stable 2.5 fallback. Check Vertex AI/AI Studio docs monthly for status.

For the "thinking_config not supported" error, yes. There may be other model-specific field restrictions (e.g., some parameters differ between Pro and Flash-Lite). Reference the feature matrix in this article.

Gemini thinking_config Not Supported on Model

Check model name before adding thinking_config. Only gemini-2.5-* models accept it. For older models (gemini-2.0-flash, gemini-1.5-pro), omit the field entirely. Best practice: use a helper that strips thinking_config when the target model is not 2.5 series.

The Error Messages You're Seeing

Thinking Support Matrix

What Actually Causes This Error

Fixes That Work (Tested Nov 2026)

1Detect Model Capability Before Config

2Strip Unsupported Fields Automatically

3Migration Path from 1.5 / 2.0 to 2.5

Preventing This Error Going Forward

Frequently Asked Questions

Related Errors

Gemini Model Migration Playbook

Check model name before adding `thinking_config`. Only `gemini-2.5-*` models accept it. For older models (`gemini-2.0-flash`, `gemini-1.5-pro`), omit the field entirely. Best practice: use a helper that strips `thinking_config` when the target model is not 2.5 series.